Govern Hugging Face models and datasets with confidence

Teams pull models and datasets straight from Hugging Face Hub, and your AI supply chain grows with every download. Cloudsmith gives you a private, fully compatible registry that proxies, caches, and governs those artifacts next to your packages and containers.

Why teams choose Cloudsmith for Hugging Face

AI adoption moves fast, and your teams drive it. Cloudsmith gives security a control point and gives developers the models they need without the wait.
Without CloudsmithDevelopers pull models straight from the public Hub, which skips scanning, policy checks, and audit trails.
With CloudsmithCloudsmith proxies and caches Hugging Face Hub, so pulls flow through your registry. Policies apply before a model reaches any workflow.
Without CloudsmithNobody can say which model version runs in production, where it came from, or what depends on it.
With CloudsmithEvery commit becomes an immutable, versioned package. You can show which model is deployed, where it came from, and reproduce any build.
Without CloudsmithModels live in a separate system from the rest of your supply chain, which leaves policy gaps and audit blind spots.
With CloudsmithModels and datasets sit with your containers and language packages, under the same access controls, policies, and audit logging.
Without CloudsmithPickle and PyTorch checkpoints carry known serialization risks, and nothing stops unsafe model file formats entering your pipeline.
With CloudsmithWrite OPA Rego policies to quarantine or block risky file formats before they reach developers.
Without CloudsmithOpen-weight models carry custom license terms, and those terms go unchecked when models are pulled directly.
With CloudsmithCloudsmith exposes license metadata to policy rules, so only approved licenses get through.
Universal format support

One registry for models, datasets, and packages. Cloudsmith is the governed store for every AI artifact and software package you use.

  • Use Hugging Face + 30 other formats
  • Push, pull, and version models and datasets with the Hugging Face SDK and CLI
  • Proxy and cache models from Hugging Face Hub with your policies applied

How we support Hugging Face

Cloudsmith is fully compatible with Hugging Face tooling, so your teams keep their workflow while you gain control over what they pull.
    Compatible with Hugging Face tooling
    Use the huggingface_hub library and the Hugging Face CLI against Cloudsmith. Set HF_ENDPOINT to your repository and your existing commands keep working.
    Upstream proxy and caching
    Configure a Cloudsmith repository as an upstream for Hugging Face Hub. Pulls flow through your registry, which caches artifacts and applies your policies first.
    Versioning by commit
    Each Hugging Face repository becomes a package and each commit an immutable version. Tags resolve to the right commit, so every build is reproducible.
    Policy on model metadata
    Write OPA Rego policies against model card metadata such as license, file formats, and training data provenance. Block risky models before they reach developers.
    Storage deduplication
    Identical files are stored once, even across repositories in the same workspace. Large model weights and datasets cost less to keep.

Get started with Hugging Face on Cloudsmith

Frequently asked questions

  1. Yes. Cloudsmith works with the huggingface_hub Python library and the Hugging Face CLI. Set HF_ENDPOINT to your Cloudsmith repository URL and HF_TOKEN to your Cloudsmith API token, and your push, pull, and download commands work as they do against the Hub.

  2. Cloudsmith supports Hugging Face model repositories and dataset repositories, including weights, configs, tokenizer files, and model cards. Hugging Face Spaces are not currently supported.

  3. No. Cloudsmith's ML Model Registry integrates with the Hugging Face SDK and CLI and stores other model files, including PyTorch, TensorFlow, and ONNX, in the same repositories as your packages and containers.

  4. Yes. Configure a Cloudsmith repository as an upstream for the public Hub. Pull requests flow through your registry, which caches artifacts and lets you apply policies before a model reaches a developer or production.

  5. A Hugging Face repository becomes a Cloudsmith package, and each commit becomes a distinct, immutable version identified by its commit hash. Tags such as main or v1.0 resolve to the correct commit.

  6. Write OPA Rego policies in Cloudsmith's Enterprise Policy Manager. Cloudsmith parses model card metadata and exposes license type, training data provenance, and file formats to your rules, so you can block risky formats like Pickle.

  7. Yes. Hugging Face repositories use native file storage deduplication. If the same files are uploaded more than once, even across repositories in a workspace, Cloudsmith stores a single copy.

  8. Set HF_TOKEN to your Cloudsmith API token before running huggingface_hub or hf commands. Cloudsmith also supports entitlement tokens for distributing models downstream, and OIDC for CI/CD such as GitHub Actions and Jenkins.

  9. Yes. Cloudsmith supports 30+ formats in one platform, so models and datasets sit beside Docker images, Python packages, npm packages, Helm charts, and Maven artifacts under the same access controls and audit logging.

  10. Yes. Upload with the huggingface_hub upload_folder function or the hf upload command, with HF_ENDPOINT pointed at Cloudsmith. If you source models from the public Hub, an upstream proxy caches them as they are pulled, so no bulk migration is needed.

  11. Yes. Set HF_ENDPOINT and HF_TOKEN in your pipeline environment and existing model download steps resolve through Cloudsmith, with caching, policy enforcement, and audit logging and no pipeline code changes.

Formats

There’s more than just Hugging Face on Cloudsmith